Live data
Assembling the frontier
Ranking the latest models and provider endpoints.
Loading model dataLive data
Ranking the latest models and provider endpoints.
Loading model dataLive model ledger
Need methodology and scenario-specific picks? Read the dedicated ai models to run locally guide.
Top overall
Muse Wonder
QI 52.4
Best value
GPT-6 Luna
$0.10/M
Fastest
Celeris-1
1848 tok/s
Largest context
Llama 4 Scout
1.3M
| # | Model | Quality | Price/M | Speed | Context |
|---|---|---|---|---|---|
| 1 | Muse WonderMeta | 52.4 | $8.0 | 113 tok/s | 1.0M |
| 2 | Muse Spark 1.3Meta | 48.1 | $2.0 | 130 tok/s | 1.0M |
| 3 | MiMo-V2.6-ProXiaomi | 46.3 | $0.54 | 178 tok/s | 1.0M |
| 4 | Qwen3.8 MaxAlibaba | 45.4 | $3.0 | 40 tok/s | 984K |
| 5 | Muse Spark 1.3 (Xhigh)Meta | 45.1 | $2.0 | 257 tok/s | 1.0M |
| 6 | GLM-5.3Z AI | 44.8 | $0.97 | 664 tok/s | 1.0M |
| 7 | Kimi K3Kimi | 43.6 | $4.3 | 444 tok/s | 1.0M |
| 8 | GLM-5.3-FlashZ AI | 41.8 | $0.12 | 1513 tok/s | 1.0M |
| 9 | Ling 3.1 FlashInclusionAI | 41.1 | $0.45 | 210 tok/s | 1.0M |
| 10 | Qwen3.8 2.4T A95BAlibaba | 39.9 | $3.0 | 156 tok/s | 1.0M |
| 11 | Qwen3.8-Flash-NextAlibaba | 39.8 | $0.23 | 51 tok/s | 1.0M |
| 12 | DeepSeek V4.1 FlashDeepSeek | 39.5 | $0.21 | 1391 tok/s | 1.0M |
| 13 | MiMo-V2.6-FlashXiaomi | 37.9 | $0.17 | 53 tok/s | 1.0M |
| 14 | DeepSeek V4 Pro 0813DeepSeek | 36 | $0.88 | 293 tok/s | 1.0M |
| 15 | DeepSeek V4 Flash VisionDeepSeek | 34.8 | $0.32 | 176 tok/s | 1.0M |
| 16 | GLM-5.3 (Low)Z AI | 34.3 | $0.97 | 607 tok/s | 1.0M |
| 17 | Qwen3.8 27B (Xhigh)Alibaba | 33.7 | Free | 360 tok/s | 1.0M |
| 18 | K2 Horizon 375B A23BInstitute of Foundation Models | 30.5 | Free | 116 tok/s | 524K |
| 19 | Kimi K3 (Low)Kimi | 30.1 | $6.0 | 124 tok/s | 1.0M |
| 20 | MiniMax-M3MiniMax | 29.2 | $0.41 | 438 tok/s | 1.0M |
| 21 | Qwen3.8 27B (Medium)Alibaba | 27.6 | $1.1 | 45 tok/s | 1.0M |
| 22 | Quasar 438BMultiverse Computing | 26.7 | $0.90 | 143 tok/s | 1.0M |
| 23 | Qwen3.8 27B (Low)Alibaba | 26.2 | $1.1 | 40 tok/s | 1.0M |
| 24 | Kimi K2.7 CodeKimi | 25.8 | $1.4 | 244 tok/s | 262K |
| 25 | Inkling SmallThinking Machines | 25.7 | $0.52 | 232 tok/s | 524K |
| 26 | Hy3Tencent | 25.3 | $0.23 | 93 tok/s | 262K |
| 27 | Qwen3.7 PlusAlibaba | 25.2 | $0.70 | 55 tok/s | 1.0M |
| 28 | Inkling (Xhigh)Thinking Machines | 25 | $1.7 | 226 tok/s | 1.0M |
| 29 | Nemotron 3 Ultra 550B A55B (Reasoning)NVIDIA | 22.9 | $0.91 | 358 tok/s | 262K |
| 30 | Ling-3.0-flash-FinInclusionAI | 22.6 | $0.11 | 281 tok/s | 262K |
Showing top 30 of 97 ranked models
View all in Explore →Decision routes
The overall score opens the conversation. These guides finish it with task-specific methodology and picks.
Ranking library
Focused rankings for the decisions engineers actually make.
Methodology
Every model is scored using the Artificial Analysis Intelligence Index — a composite of GPQA Diamond, AIME 2025, LiveCodeBench, MMLU-Pro, and other benchmarks, weighted into a single 0-100 quality score. Speed, price, and context window are tracked live across providers.
The overall ranking is a starting point. For production decisions, narrow by use case using the category pages above, then compare finalists head-to-head on Compare.
Field notes
Muse Wonder leads on overall quality right now, but the best model depends on your priorities. Coding, cost, speed, and context length all shift the answer. Use the category rankings above to find the right fit.
GPT-6 Luna currently offers one of the best quality-to-cost ratios. Open-source models on providers like Groq or Together can be even cheaper at strong quality levels.
Start with overall quality index, then narrow by what matters for your workload: cost per million tokens, output speed, context window, or a specific capability like coding or tool use. Use our Compare tool to put finalists head to head.
Celeris-1 leads on output speed right now at 1848 tokens/second. Speed matters most for real-time applications and agentic workflows with many sequential steps.
Llama 4 Scout has the biggest context window in this ranking at 1.3M. For a dedicated long-context comparison, see our largest context window page.
Data is pulled from Artificial Analysis and refreshed automatically. New models appear as soon as they have benchmark scores and provider endpoints. The ranking reflects the live state of the leaderboard.